LuckyCow/Ai_vs_real_web_images

---

license: cc0-1.0 pretty_name: AI Generated vs Real Images - Web-Scrape-3 tags: - image-classification - ai-generated - real-images - web-scraped - image

AI Generated vs Real Images

This dataset was created for binary image classification from scraped web data using the Brave search API an LLM was used to create a list of 400 diverse search queries, the brave search API was then used until the monthly free credits ran out

for each query the script first searches for that topic with the date window of Jan 1st 2010 to Dec 31st 2019, followed by the same search term with one of a rotating prefix ("Midjourney", "DALLE3", "Stable Diffusion", "AI generated") and the time window Jan 1st 2023 to Dec 31st 2025 the brave API returns 100 images for each query of which ~half get past the requirements

• image not already downloaded
• minimum 10Kb
• check image by opening it to make sure its not corrupted
• width and height must be at least 400

I used this dataset to successfully train an image classification model with 17 channels, the model was designed to look for gen AI artifacts but real images with text in them and also images of diagrams do seem to get a disproportionate amount of false positives, it is much better than my previous attempts with other datasets before i decided to make my own

Have fun <3

Structure

• real/: real images
• ai/: AI-generated images

Current local counts

• Total files: 59,872
• Real images: 31,539
• AI images: 28,333

Upload date

2026-05-02